Distributed Computing for Physics-Based Data-Driven Reduced Modeling at Scale

Presenter: Ionut Farcas, Mathematics, Computational Modeling and Data Analytics

Authors: I.Farcas, R. Gundevia, R. Munipalli, K. Willcox

Abstract: High-performance computing (HPC) has revolutionized detailed numerical simulations, with a prominent contemporary application in rotating detonation rocket engine (RDRE) design. However, these RDRE simulations are computationally demanding, requiring millions of core hours, making them impractical for engineering tasks like design exploration, risk assessment, and uncertainty quantification. Reduced-order models (ROMs) offer a solution by constructing fast yet sufficiently accurate approximations that serve as surrogates for the high-fidelity model. We present a distributed memory algorithm for the fast and scalable construction of predictive physics-based ROMs. This algorithm is trained from sparse datasets of extremely large state dimension, learning structured physics-based ROMs that approximate the underlying dynamical systems. We demonstrate the algorithm's scalability using up to 2,048 cores on the Frontera supercomputer at the Texas Advanced Computing Center. Our focus is a real-world RDRE simulation, where one millisecond of physical time requires one million core hours. Using a training dataset of 2,536 snapshots, each with a state dimension of 76 million, our distributed algorithm constructs a predictive reduced model in just 13 seconds on 2,048 cores.